Data centers used 4.4 percent of American electricity in 2023. Nobody measured it.

One hundred and seventy-six terawatt-hours.

That is the number. It appears in utility filings, in state legislative testimony, in investor decks, and in most of the journalism written about artificial intelligence and the power grid over the past eighteen months. It comes from a single document: the 2024 United States Data Center Energy Usage Report, published by Lawrence Berkeley National Laboratory in December 2024 under report number LBNL-2001637, prepared for the Department of Energy because the Energy Act of 2020 told Congress it wanted an update to a study last done in 2016.

The report is careful, long, and honest about itself. It is also a model. No meter anywhere produced the figure 176, and the authors say as much in plain language on page seven: "The lack of direct energy data available in a sector with rapidly evolving technologies limits the analysis in this report."

That sentence is the story.

How much electricity do data centers use in the United States?

US data centers consumed about 176 terawatt-hours in 2023, or 4.4 percent of total national electricity consumption, according to Lawrence Berkeley National Laboratory. The figure covers servers, storage, network equipment and the cooling and power infrastructure that supports them. It is an estimate built from equipment shipment data, not a measurement.

The arithmetic checks. Total US electricity end use in 2023 was 4,011,171,517 megawatt-hours according to the Electric Power Annual, Table 2.2, which the Energy Information Administration publishes as a spreadsheet. Divide 176 terawatt-hours by 4,011 and the result is 4.39 percent. The report's own earlier waypoint ties out the same way: about 76 terawatt-hours in 2018 against 4,003 terawatt-hours of total end use gives 1.90 percent, and the report says 1.9 percent.

So the denominators are solid. The denominators come from utilities, which are required to report, and the reporting has been going on for decades. It is the numerator that is inferred.

Where does the 176 figure actually come from?

The estimate is bottom-up. Berkeley counted the installed base of servers, storage drives and network ports from shipment data, assumed a power draw and an operating pattern for each class of equipment, applied an assumed power usage effectiveness for the buildings, and summed the result. Every input is a judgment about hardware nobody outside the operators can inspect.

There is a defensible reason for choosing that method over the obvious alternative. Utilities publish interconnection queues, and those queues have become the raw material for a great many alarming forecasts. The problem is that a developer can and does file for interconnection in five places for one facility, so the queue counts the same data center several times and counts projects that will never be built. Berkeley notes that the bottom-up approach "avoids overestimation that can be caused by tracking data center load for projects that have not yet selected a power provider."

Trading one error for another is a reasonable trade. It is still a trade. The bottom-up method is only as good as the shipment data and the wattage assumptions underneath it, and both are supplied largely by proprietary market analysts rather than by anyone under an obligation to be right.

Why does the government not simply measure it?

It is starting to. On 25 March 2026 the Energy Information Administration announced three voluntary pilot field studies: web-based surveys in Texas and Washington state, and in-person interviews in Northern Virginia and Washington, DC. The agency identified 196 companies operating data centers in those regions. Each will be asked to report on at least one facility.

Read that again with the dates attached. The estimate everyone quotes was published in December 2024 and its newest historical year is 2023. The first federal attempt to collect the underlying data directly began in March 2026, is voluntary, is a pilot, and covers three regions.

The agency is not hiding the gap. "A tremendous amount of excellent work goes into our retrospective consumption surveys, but they were conceived decades ago," EIA Administrator Tristan Abbey said in the release. The questionnaire will ask about energy sources, electricity consumption, site characteristics, server metrics and cooling systems, which is precisely the list of things the Berkeley model has to assume.

None of this makes 176 wrong. It makes it provisional in a way the number's public career does not reflect.

How fast has data center electricity use actually grown?

The trajectory in the Berkeley report is the least contested part of it. Data center electricity use held near 60 terawatt-hours from 2014 through 2016, rose to about 76 terawatt-hours by 2018, and reached 176 terawatt-hours in 2023. The compound growth rate was roughly 7 percent from 2014 to 2018 and 18 percent from 2018 to 2023.

The flat decade is the interesting half. Through the early and mid 2010s the industry grew enormously while its electricity use barely moved, because workloads migrated from company server closets into hyperscale and colocation facilities that ran far more efficiently. Berkeley measures that shift directly: infrastructure, meaning cooling and power conversion rather than computing, accounted for 40 percent of total data center electricity in 2014 and 30 percent in 2023.

Efficiency won for ten years. Then it stopped winning, and the reason was accelerated servers built for artificial intelligence entering the installed base from 2017 onward.

What do the measured electricity numbers show?

The national data does not break out data centers, so it cannot confirm the Berkeley figure. It can be read for the shape of the thing. Standalone data centers land in the commercial sector, and the commercial sector has been behaving unusually.

From Electric Power Monthly Table 5.1, retail electricity sales to all sectors rose from 3,811,150 gigawatt-hours in 2019 to 4,058,007 in 2025, a gain of 246,857 gigawatt-hours. Commercial sales supplied 132,609 gigawatt-hours of that, which is 53.7 percent of all the growth in American electricity sales over six years, from a sector that was 35.7 percent of the 2019 base. Residential added 74,704, industrial added 39,864, and transportation fell slightly.

Narrow the window and the pattern holds. Between 2023 and 2025 total sales grew 183,754 gigawatt-hours and commercial supplied 85,377 of it, or 46.5 percent.

Now put the two datasets against each other, which neither publisher does. Berkeley's own scenario range implies data center electricity use growing between 13 and 27 percent a year after 2023. Start at 176 terawatt-hours and apply those rates for two years and data centers would have added between 49 and 108 terawatt-hours by 2025. Against the 184 terawatt-hours of measured growth in total US retail sales over the same two years, that is somewhere between 27 percent and 59 percent of everything.

The honest reading of that calculation is that it is consistent with data centers being the largest single driver of American electricity demand growth, and that it cannot distinguish between a large role and an overwhelming one. A range that runs from a quarter to three fifths is not a finding. It is the shape of what is currently knowable.

How wide is the uncertainty about 2028?

Berkeley declined to publish a point estimate for 2028 and gave a range instead: 325 to 580 terawatt-hours, or 6.7 to 12.0 percent of forecast national consumption, which at an assumed 50 percent capacity utilization works out to 74 to 132 gigawatts of power demand.

The width deserves more attention than it gets. The spread between the ends is 255 terawatt-hours. That is larger than the entire 2023 estimate of 176. It is also larger than the 247 terawatt-hours by which total US retail electricity sales grew across the whole of 2019 to 2025. The uncertainty band around one sector's five-year forecast exceeds six years of measured growth in the national total.

What drives the spread is not exotic. It is how many GPUs get shipped, how hard they are worked once installed, and which cooling systems the buildings choose. Nobody outside the operators knows any of the three.

Do these estimates get revised?

They do, and the record is instructive. Berkeley's 2016 report put 2014 data center consumption at about 70 billion kilowatt-hours, or 1.8 percent of national electricity use. The 2024 report, rerunning the method on better data, puts 2014 nearer 60 terawatt-hours. The best available federal estimate of a year that had already happened moved by roughly a sixth.

The report explains why, at some length, and the explanations are technical rather than embarrassing. Shipment data was revised down slightly. Space-type definitions changed. More granular analyst data showed more servers sitting in small internal facilities, running at low utilization, than had been assumed. And the categorization of servers moved from price point to processor count, which retired an old assumption that had extrapolated wattage growth for high-end machines out to figures that, in the report's words, "are not in line with any servers we have seen produced and shipped since then."

That is what a good laboratory does with a number it no longer believes. It is also a warning about how much confidence any single vintage of this estimate can carry.

What does EIA now project for the long run?

For the Annual Energy Outlook 2026, released 8 April 2026, the agency updated its Commercial Demand Model to report data center server electricity separately from the broader commercial computing category for the first time. On the agency's slide, data center load "is emerging as the dominant driver of long-term U.S. electricity growth."

The projections, published in May 2026, put servers alone at an estimated 7 percent of commercial sector electricity consumption in 2025, rising to between 22 and 33 percent of commercial building electricity by 2050. In absolute terms server consumption reaches 446 to 818 billion kilowatt-hours in 2050 depending on the case. The agency assumes data center floorspace needs as much as 2.9 times the cooling energy of ordinary commercial floorspace, and that server load is essentially flat across every hour of the day, which is a meaningful thing for a grid built around evening peaks.

One assumption in the baseline case is worth naming, because it does a great deal of work. After 2040 the Counterfactual Baseline assumes servers get 10 percent more efficient in average operational power draw every three years, over and above historical efficiency trends. The High Electricity Demand case makes no such assumption. Much of the gap between 446 and 818 is that single choice, and the AEO in most cases reflects only laws and regulations as of December 2025.

What comes with the electricity?

Two figures from the Berkeley report cut against the usual framing. Data center electricity in 2023 carried an estimated indirect water footprint of nearly 800 billion liters and 61 billion kilograms of carbon dioxide equivalent, which sounds enormous and is. But per unit consumed, data center electricity was slightly more water intensive than the national average and slightly less carbon intensive: 4.52 liters per kilowatt-hour against 4.35, and 0.34 kilograms against 0.35.

Direct water use, the water evaporated in the buildings themselves, went from 21.2 billion liters in 2014 to 66 billion in 2023. Hyperscale and colocation facilities accounted for 84 percent of that 2023 total, and small internal data centers had fallen to 12 percent.

The carbon intensity finding follows from geography rather than virtue. Data centers cluster in places whose grid mix happens to be somewhat cleaner than average, and the report is explicit that its method cannot see power purchase agreements or behind-the-meter generation at individual sites, which could move the answer in either direction.

Frequently asked questions

Is 176 terawatt-hours the current number? It is the most recent federal historical estimate, and it is for 2023. No official figure for 2024 or 2025 exists. Anything presented as a current data center consumption number is either an extrapolation from the Berkeley baseline or a projection from a scenario.

Does 4.4 percent include cryptocurrency mining? No. Berkeley treats Bitcoin mining separately, in its own chapter, with its own estimates and its own scenario ranges out to 2028.

Is US electricity demand actually rising? Yes, and it is the first sustained rise in a long time. EIA's January 2026 Short-Term Energy Outlook forecast electricity use growing 1 percent in 2026 and 3 percent in 2027, which would be the first four consecutive years of growth since 2007 and the strongest four-year stretch since 2000. The agency attributes the surge to large computing facilities.

When will there be measured data? Unknown. The EIA pilot launched in March 2026 covers 196 companies across three regions and is voluntary. A pilot is a test of feasibility, not a data series, and the agency has not published a timetable for turning it into one.

What should be treated as solid? The national totals, which come from utility reporting. The sector split, on the same basis. The direction and rough magnitude of data center growth, which every method agrees on. What should not be treated as solid is any single figure past 2023, or any projection quoted without the range it came in.